Xirang Kaipu, a base-model company for physical interaction, has closed consecutive seed and angel rounds worth several hundred million yuan and now carries a valuation of 500 million dollars. The round was led by Dunhong Capital, with Huakong Fund, Sanhua Holdings, Yinxinggu Capital, Zeran Capital, Benjian Fund, Angel Cornerstone and Biaopu Investment following.
The capital funds large-scale pretraining of a native physical-AI model, construction of real physical-interaction data, recruitment of core research talent and scenario validation.
Xirang Kaipu registered in July 2026 and was co-founded by two core figures from Huawei’s large-model system. Founder and chief executive Li Yin was Huawei Cloud’s large-model chief technology officer and head of its industry large-model development department, a core member from the ground up of Huawei’s model work, long leading base models, multimodal systems, video generation, scientific computing and industry models. Co-founder and chief scientist Zhang Hanwang is Huawei’s multimodal chief scientist, known for causal machine learning and multimodal intelligence. Causal learning pushes AI past correlation toward causality, the theoretical base for truly grasping physical laws, and he has published 228 papers at the top tier.
The pairing means the team can both train a base model from the bottom and land it in industry. Its strategy skips robot bodies and single-point applications to build a physical base model, the Large Physics Model, that understands physical laws, predicts state evolution, plans long-horizon actions and supports closed-loop execution.
A base built on the Bellman equation
Xirang Kaipu roots the model in the Bellman-equation optimum, breaking a long-horizon decision into step-by-step small choices so the combinatorial explosion resolves mathematically. In architecture the Large Physics Model decouples world inference from video rendering: a unified autoregressive structure handles causal history and infers what happens, while a diffusion model renders how it looks.
Around the model the firm built three cooperating platforms. The data platform compiles physical-world experience into trainable assets, running multimodal access, cleaning, labelling, quality checks and version control with cross-entity, cross-model neutrality. The augmentation platform lets base-model ability grow into a customer scene cheaply, offering an augment, deploy, iterate toolchain so a client needs no large algorithm team to fine-tune a base model into its own industry model. The evaluation platform makes capability verifiable and faults locatable, automating scenario build, strategy execution and reporting so failed samples are found, supplemented and re-verified, then fed back into the next training round. Data feeds, augmentation forms, evaluation finds faults, failed samples return, and each loop strengthens the model.
From robots to the wider physical world
Commercially the plan runs in two steps. First, close the robot-scene loop with body makers and scenario partners to form a verifiable delivery sample. Second, replicate the same base across the broader physical world, from autonomous-driving closed-loop simulation and extreme-weather prediction to drug-molecule generation, all problems at heart about understanding physical laws and predicting state evolution, the Large Physics Model’s home ground.
The firm notes that in the smartphone era hardware makers built devices, operating systems built the base and developers grew the ecosystem, and it will not chase every scene but deepen the base so partners use it in their own. A 500 million dollar valuation two months after registration is a stark bet that the platform layer, not the robot, is where physical AI will be won.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/50640d4b55d5cba175fb84f15d679f19/robot/news/70473151-seeds-%E5%8D%8E%E4%B8%BA%E5%A4%A7%E6%A8%A1%E5%9E%8B%E6%A0%B8%E5%BF%83%E6%88%90%E5%91%98%E5%88%9B%E4%B8%9A/.